# Artificial intelligence-based diagnostic model for schizophrenia in individuals living with HIV

## TL;DR

This study develops an AI model to diagnose schizophrenia in people living with HIV using blood parameters, achieving high accuracy and identifying key biomarkers.

## Contribution

The novel contribution is the development and validation of an AI-based diagnostic model for schizophrenia in HIV-positive individuals using hematological data.

## Key findings

- The Lasso regression model achieved high diagnostic performance with an AUC of 0.966 and accuracy of 0.897.
- PDW, MPV, and MCV were identified as the most influential hematological biomarkers for schizophrenia diagnosis in HIV patients.
- Machine learning models using routine blood parameters showed strong potential for schizophrenia detection in PLWH.

## Abstract

Schizophrenia is one of the most prevalent severe mental disorders among people living with HIV (PLWH). Delayed diagnosis and misdiagnosis contribute to poor prognosis and substantial economic burden in this population. However, there are currently no validated diagnostic models available for schizophrenia in PLWH.

PLWH attending annual follow-ups at Yunnan Provincial Hospital of Infectious Diseases/Yunnan AIDS Care Center were enrolled. Hematological parameters were compared between PLWH with schizophrenia (HIV-Scz) and those without (HIV-non-Scz) and diagnostic models were constructed using six machine learning algorithms. Model performance was evaluated comprehensively using area under the curve (AUC), accuracy, F1 score, recall, precision, and decision curve analysis. SHapley Additive exPlanations (SHAP) were applied to determine the relative importance of each feature.

A total of 186 participants were included in this study, including 62 with clinically confirmed schizophrenia who were receiving antipsychotic treatment at the time of blood sampling. Compared with the HIV-non-Scz group, the HIV-Scz group exhibited significant differences across multiple hematological parameters. Six machine learning models constructed using 28 routine blood parameters demonstrated diagnostic capability, among which the Lasso regression model achieved the best overall performance, with mean AUC (0.966 ± 0.016), F1-score (0.839 ± 0.067), and accuracy (0.897 ± 0.037), together with favorable precision (0.867 ± 0.061) and recall (0.821 ± 0.111). Decision curve analysis indicated that this model provided a higher net benefit within clinically relevant threshold probability ranges. Furthermore, SHAP analysis identified PDW, MPV and MCV as the most influential features contributing to model predictions.

Routine hematological parameters may serve as potential diagnostic biomarkers for schizophrenia in PLWH, although medication-related effects in treated patients cannot be excluded.